The official experiment cohort contains 880 recordings / 440 matched pairs. Each pair contains one collision and one near miss, kept in the same split. The supplied split manifests define the fixed evaluation cohort.

Scenario	Train videos	Val videos	Test videos	Total
Head-on	66	26	26	118
Close turning	86	24	30	140
Following/braking	52	42	44	138
Pedestrian walking past	66	32	32	130
Pedestrian walking then stopping	8	12	12	32
Unoccluded pedestrian running	28	8	6	42
Stationary car, plain background	50	24	26	100
Stationary car, parked background cars	36	22	22	80
Stationary pedestrian	48	30	22	100
Total	440	220	220	880

Structure

metadata.jsonl                         # Complete catalog, one row per recording.
splits/{train,val,test}.jsonl           # Same rows, separated by split.
release_membership.json                # IDs used in the original experiments.
data/{train,val,test}/PAIR_ID/
  pair_config.json                     # Native generation and crop settings.
  collision/                           # near_miss/ has the same structure.
    metadata.json                      # Portable metadata and asset checksums.
    events.npy                         # Detector input; unchanged source bytes.
    capture_metadata.json              # Native capture details and trajectories.
    source_metadata.json               # Original metadata, preserved verbatim.
    trajectory.json                    # Aligned 400 ego/target poses.
    review.mp4                         # Human preview, not a detector input.

Some stationary-pedestrian previews show both arms in one pair-level review.mp4; use assets.review_video.path and preview_shared_between_arms in the catalog. Metadata uses split: "validation" for the val/ folder. Other fields include scenario, target type, static, dense_traffic, town/roads, speeds and background. Unknown values remain null. FORMAT.md explains clocks and encoding.

The dataset viewer reads viewer/{train,val,test}.parquet, which contains the same 880 catalog rows with a shared column schema. The original JSON catalogs and all recording files remain unchanged. Additional null fields in the viewer only represent metadata absent from a particular source.

Download and verify

Open the anonymous review repository. Reviewers can browse and download files without an account or access token. Use the file browser for individual files or Download All for the complete release. Extract the download into a folder of your choice, then verify it:

python -m pip install 'numpy>=1.26,<3'
cd /path/to/paired-data
python verify_release.py --deep-events

Replace /path/to/paired-data with the extracted folder containing verify_release.py. Full verification requires the complete release and checks for missing or corrupted files. A partial download cannot pass this check. MP4 files are inspection previews and are not detector inputs.

Scope

The split is frozen at 50/25/25, with 20 explicit background-OOD pairs confined to validation/test. Two generation pipelines contribute 600 and 280 recordings. Selection retained all moving-target pairs and reduced only the stationary subsets. Earlier excluded recordings, occluded-runner trials, original CARLA corpus data and paper-only illustrations are not this cohort.

Capture audits and source qualifications are preserved. In particular, the stationary-pedestrian source still says manual_review: pending and scientific_admission: false; packaging does not upgrade that status. Integrity checks establish matching files, labels and timing, not freedom from dataset bias. Raw event streams and duplicate dense-frame caches are omitted. A dataset license has not yet been specified.

Keep both outcomes of each pair in the supplied split. Preserve the catalog, capture metadata and checksums when copying the data; do not rebuild the catalog from a partial download.

Downloads last month: 151

Total file size:

40.7 GB